InfraRSI, an LLM distributed learning optimization agent, achieves a geometric mean speedup of 1.83x

• Proposed InfraRSI, an agent harness that jointly explores execution settings (knobs) and source code for LLM distributed learning, and reflects changes after contract-based validation
• Achieved a geometric mean speedup of 1.83x and increased average MFU from 17.4% to 30.0% across four workloads on H100/H200
• The gap compared to the baseline (RE-Bench 1.62x) was small for the dense model (Qwen3.5-9B), whereas the difference was significant for two MoE models, with InfraRSI (1.51x, 2.75x) outperforming the baseline (1.17x~1.33x)
• Only executions passing 16 invariant conditions are considered valid, ensuring reliability by rejecting 33.0% of the 1,212 stored entries

Optimization Strategy and Validation

• Retried OOM failures by combining them with memory-saving settings using the Budgeted Compositional Search (bcomp) strategy
• Corrected LLM speed prediction errors using empirical Bayes shrinkage and reverse-direction probing
• Achieved a 1.82x improvement in cross-SKU transfer through an experience bank that accounts for hardware-specific effect differences

Limitations and Conditions

• Validated only on H100/H200 at a scale of 8~16 nodes; no results for other GPUs/clusters
• Final model quality is out of scope for inspection, and the LLM used for exploration is undisclosed